Transfer Learning Approaches for Brain Metastases Screenings.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Brain metastases detection on MRI is crucial for patient outcomes.
• Automatic segmentation aids in preventive exams and remote diagnostics.
• Deep learning models offer potential for enhanced segmentation accuracy.

Purpose of the Study:

• To evaluate the effectiveness of transfer learning for brain metastasis segmentation on MRI.
• To compare transfer learning models with models trained from scratch.
• To assess the impact of custom loss functions on segmentation performance.

Main Methods:

• Trained three deep learning models using transfer learning on public and private datasets.
• Fine-tuned models on a smaller, specialized dataset.
• Compared performance against models trained from scratch.
• Utilized a custom Tversky and Binary Cross-Entropy loss function.

Main Results:

• Transfer learning models outperformed scratch-trained models, though not statistically significant.
• The custom loss function effectively managed class imbalance and reduced false negatives.
• Medical experts observed good performance on larger tumors but limitations with smaller, scattered ones.

Conclusions:

• Transfer learning and custom loss functions show potential for medical imaging segmentation.
• Model limitations exist in detecting very small tumors in complex cases.
• Further research is needed to address challenges in segmenting subtle lesions.